genmake_engine.py

import os
import time
import numpy as np
from PIL import Image
import torch
from transformers import pipeline
import cv2
from rembg import remove, new_session

from photo_prep import studio_enhance_image
from glb_export import export_pointcloud_glb
from genmake_voxel_engine import build_importance_map, build_voxel_volume, sample_points_from_volume

PRODUCTS = {
    "keychain_rectangle_20x30x15": {"size": (17.0, 25.5, 12.0)},
    "keychain_square_30x30x10": {"size": (25.2, 24.48, 7.82)},
    "keychain_heart_30x30x12": {"size": (25.0, 25.0, 9.0)},
    "cube_50x50x50": {"size": (42.0, 40.80, 39.10)},
    "cube_60x60x60": {"size": (51.0, 51.0, 48.0)},
    "cube_80x80x80": {"size": (68.0, 68.0, 65.0)},
    "block_50x50x80_portrait": {"size": (42.0, 68.0, 40.80)}, 
    "block_60x60x100_portrait": {"size": (51.0, 85.0, 51.0)},
    "block_80x80x120_portrait": {"size": (68.0, 102.0, 68.0)},
    "block_80x50x50_landscape": {"size": (68.0, 42.0, 40.80)},
    "block_100x60x60_landscape": {"size": (85.0, 51.0, 51.0)},
    "plaque_100x100x20": {"size": (85.0, 85.0, 15.0)},
    "plaque_150x100x30_landscape": {"size": (125.0, 85.0, 24.0)},
    "plaque_100x150x30_portrait": {"size": (85.0, 125.0, 24.0)},
}

QUALITIES = {
    "fast": {"points": 900000, "voxel": 0.020},
    "balanced": {"points": 1500000, "voxel": 0.015},
    "premium": {"points": 2200000, "voxel": 0.012},
}

_DEPTH_MODEL = None
_REMBG_SESSION = None

def _ts():
    return time.strftime("%Y%m%d_%H%M%S")

def _ensure(p: str):
    os.makedirs(p, exist_ok=True)

def _pil_to_bgr(pil_image):
    rgb = np.array(pil_image.convert("RGB"))
    return cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)

def _resize_for_ai(pil_image: Image.Image, max_size=1024) -> Image.Image:
    w, h = pil_image.size
    if max(w, h) > max_size:
        scale = max_size / float(max(w, h))
        pil_image = pil_image.resize((int(w * scale), int(h * scale)), Image.Resampling.LANCZOS)
    return pil_image

def _generate_shape_mask(w, h, shape_type, scale, offset_x_pct, offset_y_pct):
    mask = np.zeros((h, w), dtype=np.float32)
    cx = (w // 2) + int((offset_x_pct / 100.0) * w)
    cy = (h // 2) + int((offset_y_pct / 100.0) * h)
    
    if shape_type == "Round":
        r = int(min(w, h) * 0.48 * scale)
        cv2.circle(mask, (cx, cy), r, 1.0, -1)
    elif shape_type == "Square":
        s = int(min(w, h) * 0.95 * scale)
        x1, y1 = cx - s//2, cy - s//2
        x2, y2 = cx + s//2, cy + s//2
        cv2.rectangle(mask, (x1, y1), (x2, y2), 1.0, -1)
    elif shape_type == "Heart":
        t = np.linspace(0, 2 * np.pi, 100)
        x = 16 * np.sin(t)**3
        y = 13 * np.cos(t) - 5 * np.cos(2*t) - 2 * np.cos(3*t) - np.cos(4*t)
        x = (x / 16.0) * (min(w, h) * 0.45 * scale)
        y = -(y / 16.0) * (min(w, h) * 0.45 * scale) 
        pts = np.vstack((x + cx, y + cy)).astype(np.int32).T
        cv2.fillPoly(mask, [pts], 1.0)
    elif shape_type == "Star":
        pts = []
        outer_r = min(w, h) * 0.48 * scale
        inner_r = outer_r * 0.4
        for i in range(10):
            angle = i * np.pi / 5 - np.pi / 2
            r = outer_r if i % 2 == 0 else inner_r
            pts.append([cx + r * np.cos(angle), cy + r * np.sin(angle)])
        pts = np.array(pts, dtype=np.int32)
        cv2.fillPoly(mask, [pts], 1.0)
    else:
        mask = np.ones((h, w), dtype=np.float32)
        
    return mask

def _make_depth01(pil_image: Image.Image, device: str, fp16: bool) -> np.ndarray:
    global _DEPTH_MODEL
    if _DEPTH_MODEL is None:
        print("⏳ Loading Depth AI into GPU...")
        _DEPTH_MODEL = pipeline("depth-estimation", model="Intel/dpt-hybrid-midas", device=0 if device.startswith("cuda") else -1, torch_dtype=torch.float16 if (fp16 and device.startswith("cuda")) else torch.float32)
    print("⚙️ Running Depth Inference...")
    out = _DEPTH_MODEL(pil_image)
    d = np.array(out["depth"]).astype(np.float32)
    mn, mx = float(d.min()), float(d.max())
    if mx - mn < 1e-8: return np.zeros_like(d, dtype=np.float32)
    return (d - mn) / (mx - mn)

def generate_crystal(image_path, product, face_mode, bg_mode, shape_mask, mask_scale, mask_x, mask_y, diorama_mode, quality, use_enhanced, fp16, detail_boost, depth_vol, out_root="GenMake_Designs"):
    global _REMBG_SESSION
    print(f"\n🚀 STARTING GENERATION FOR: {product}")
    start_time = time.time()
    
    if product not in PRODUCTS: product = "block_50x50x80_portrait"
    sx, sy, sz = PRODUCTS[product]["size"]
    
    if quality not in QUALITIES: quality = "balanced"
    base_points = QUALITIES[quality]["points"]
    voxel_mm = QUALITIES[quality]["voxel"]

    max_area = 42.0 * 68.0
    area_ratio = (sx * sy) / max_area
    dynamic_total_points = int(base_points * area_ratio)

    print("📸 Loading Image...")
    raw_pil = Image.open(image_path).convert("RGB")
    if use_enhanced: raw_pil = studio_enhance_image(raw_pil)
    pil = _resize_for_ai(raw_pil, max_size=1024)
    bgr = _pil_to_bgr(pil)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    depth01 = _make_depth01(pil, device=device, fp16=fp16)
    
    # --- UPGRADE 1: Depth Confidence Laplacian Filter ---
    print("🧹 Applying Depth Confidence Filter...")
    depth_grad = cv2.Laplacian(depth01, cv2.CV_32F)
    depth01[np.abs(depth_grad) < 0.002] *= 0.8

    print("✂️ Masking...")
    if bg_mode != "off":
        if _REMBG_SESSION is None: _REMBG_SESSION = new_session("u2net_human_seg") 
        bg_removed = remove(pil, session=_REMBG_SESSION)
        fg_mask = (np.array(bg_removed)[:, :, 3] > 128).astype(np.float32)
        fg_mask = cv2.GaussianBlur(fg_mask, (5, 5), 0)
    else:
        fg_mask = np.ones_like(depth01, np.float32)

    if shape_mask not in ["None", "Rectangle"]:
        print(f"✂️ Slicing 3D Volume to Shape: {shape_mask}...")
        geom = _generate_shape_mask(pil.size[0], pil.size[1], shape_mask, mask_scale, mask_x, mask_y)
        fg_mask = fg_mask * geom

    if not diorama_mode:
        depth01[fg_mask < 0.5] = 0.0

    valid_depths = depth01[fg_mask > 0.5]
    if len(valid_depths) > 0:
        d_min, d_max = float(valid_depths.min()), float(valid_depths.max())
        depth01 = np.clip((depth01 - d_min) / (d_max - d_min + 1e-8), 0.0, 1.0)

    base = os.path.splitext(os.path.basename(image_path))[0]
    out_dir = os.path.join(out_root, f"{base}_{_ts()}", "engine")
    _ensure(out_dir)

    depth_png = os.path.join(out_dir, "GenMake_Depth.png")
    weight_png = os.path.join(out_dir, "GenMake_WeightPreview.png")
    glb_path = os.path.join(out_dir, "GenMake_EngraveReady.glb")

    Image.fromarray((depth01 * 255).astype(np.uint8)).save(depth_png)
    
    print("🧠 Building HDR Maps...")
    imp01 = build_importance_map(bgr, depth01, fg_mask)
    Image.fromarray((imp01 * 255).astype(np.uint8)).save(weight_png)
    
    V = build_voxel_volume(depth01, imp01, fg_mask, z_layers=60, thickness_layers=6)
    
    print(f"⚡ Processing 3D Shell...")
    points = sample_points_from_volume(V, dynamic_total_points, voxel_mm, sx, sy, sz, detail_boost, depth_vol, diorama_mode=diorama_mode, seed=123)

    print("💾 Exporting .GLB...")
    export_pointcloud_glb(points, glb_path)

    elapsed = time.time() - start_time
    print(f"✅ DONE in {elapsed:.2f}s.")
    
    log_text = f"✅ SUCCESS: {product}\nShape Output: {shape_mask}\nTotal Points: {len(points):,}\nTime: {elapsed:.2f}s"
    return log_text, [glb_path, depth_png, weight_png], depth_png, weight_png